How and Whether Large Language Models Can Help Social Scientists

Applications of GPT (or other LLMs) in social science:

• Nov 11: Text classification, scaling and topic modelling
• Feb 21: Simulate survey responses for counterfactual persons
• Mar 7: Generate persuasive political arguments
• Mar 22: Ideological scaling of US senators
• Mar 27: Out-perform crowd workers for “manual” coding

Motivation

• Hard to keep up; hard to know where to start
• (I argue) In some cases, confusion over technology has already led to misapplication

This Talk

• Technical explainer of GPT
  • At a level that helps understand what it is and why it behaves as it does.
• Discussion of current applications
  • Innovations
  • Shortcomings
  • Guidelines
• Brief speculation on where this is headed

Visual Demo

Text In, Text Out

How do you “Model” Language?

• We are familiar with modelling numerical processes (i.e. regression)
• How do you construct a model that goes from language to language?

Language as a Sequence

“GPT for Social Research”

• As a sequence of words:
  • (GPT, for, Social, Research)
  • \(S_1=(w_1, w_2, w_3, w_4)\)
• Actually we don’t use words (will come back to this)

Sequence-to-Sequence

Model to map from one sequence to another:

• \(M(S_1) \rightarrow S_2\)
• Conversation Model: “How are you?” → “I am great!”
• Translation Model: “How are you?” → “¿Cómo estás?

Simplifying the Problem

Start with the word “Once”:

• What words could come next?

Language as Conditional Probabilities

What words could come next?

• \(Pr(\text{you} | \text{Once}) = 0.5\)
• \(Pr(\text{upon} | \text{Once}) = 0.2\)
• …

Once
├── you  (0.5)
├── upon (0.2)
└── [...]

Following one branch:

• \(Pr(\text{are} | \text{Once you}) = 0.21\)
• \(Pr(\text{finish} | \text{Once you}) = 0.01\)

Once
├── you
│   ├── are     (0.21)
│   ├── finish  (0.01)
│   └── [...]
└── upon

Following the other branch:

• \(Pr(\text{a} | \text{Once upon}) = 0.99\)
• \(Pr(\text{time} | \text{Once, upon, a}) = 0.99\)

Once
├── you
│   └── [...]
└── upon
    └── a (0.99)
        └── time (0.99)

Autoregressive Language Models

• Input: “Once upon”
• Step 1: M(“Once upon”) → “a”
• Step 2: M(“Once upon a”) → “time”
• Step 3: M(“Once upon a time”) → “,”
• Step 4: M(“Once upon a time,”) → “there”
• […]
• Up to some maximum window size!

Demo with davinci

Tokenization and Vocabularies

• Space of all words is very large
• Individual characters carry too little signal about what comes next
• Sub-word tokenization: something in between
• GPT-2 has a vocabulary size of 50,257 unique tokens

Tokenization Visualized:

So, what is GPT?

• GPT(-2, 3, 3.5) are a collection of sequence-to-sequence models that use auto-regressive language generation to produce textual outputs from textual inputs.
• Internally, they treat all of language as a conditional probability distribution over tokens.
  • Information is stored/retrieved as the most likely continuation of an input.
  • With caveats about “most likely” (to be discussed)
• How is this probability distribution learned?

Training

How to train a GPT

• The answer is surprisingly simple:
• Next word prediction
• … a lot of parameters
• … and a lot of examples

Next Word Prediction

A Lot of Parameters

• \(Y = \beta_2X_2 + \beta_1X_1 + \beta_0\) has 3 parameters
• GPT-3 has approx. 175,000,000,000 parameters!

Parameter Inflation

A Lot of Examples

• Approx. “300 billion training tokens, \(3.14E+23\) FLOPS” (Brown et al. 2020, Appendix D)

Where do these examples come from?

• CommonCrawl (filtered)
  • 41 months (2016-2019) of crawled Internet content
  • Deduplicated and filtered from 45TB to 570GB.
• WebText2: OpenAI’s internal dataset.
  • Starting point all outbound links from Reddit with at least 3 karma:
  • “heuristic indicating whether people found something interesting, educational or funny.” (Brown et al. 2020)
• Books1 and Books2: bookcorpus and a mystery
• English-language Wikipedia

As a Completion Tool?

• Some uses for a most-likely-continuation tool of the Internet
  • Creative writing?

Is Completion… Everything?

• As size of models increased, a surprising behavior emerged:
• GPT-3 could do tasks that it had not been trained on, without further training
• Called “In-Context Learning”

What does this look like?

One-Shot

Few-Shot

Evaluation Example

LAMBADA:

• “Alice was friends with Bob. Alice went to visit her friend ____.”
• “George bought some baseball equipment, a ball, a glove and a ____.”

More Parameters = Better Zero/Few-Shot

Foundation Models (Bommasani et al. 2021)

• Pivot away from task-specific models and architectures
• towards one-model-fits-all approaches.

Instruction-Tuning (Ouyang et al 2022)

• Subsequent model (InstructGPT) added novel training approach:
  • Reinforcement Learning with Human Feedback (RLHF)
• Objectives:
  • Make model better at following instructions (instead of continuing)
  • Make model less likely to produce harmful/toxic language.

Completion ≠ Instruction

Toxicity

• Model trained to give most likely completion from Internet data
• Have you been on Reddit?

Example of Non-Tuned Output (from OPT-30B)

CW: Obscenities

Instruction-Tuning: Process

Who are the Annotators?

• 40 Upwork/Scale AI contractors
• “group of labelers who were sensitive to preferences of different demographic groups” (Ouyang et al. 2022)
• Screened using test on ability to filter toxic content

Recap: what is GPT now?

• GPT-2: Start with model that gives most likely continuation of sequence.
• GPT-3: Make it bigger. Gains zero-shot abilities.
• InstructGPT: Adjust model to give best response to instruction.
• ChatGPT: unclear exactly what they changed (only a short blog post from OpenAI).
  • Speculation: new user interface, more RLHF, add special tokens to structure dialogue.

Innovation 1: GPT as a Coder

Innovation 1: GPT Outperforms Crowd Coding

Challenge: Unknown Estimator Properties

• Predictions given by GPT are:
  • biased (socially and statistically) in an unknown way
  • sensitive to exact phrasing of prompt
• Problem: we don’t know if/when it will fail–or if it has failed!
• Solution: forthcoming work

Innovation 2: GPT as a Respondent

Silicon Sampling (Argyle et al. 2023): prompt model with demographic traits then recover response:

A Few Warnings

• Technical: In-context learning ≠ Conditioning
  • GPT always returns \(P(S_{out} | \mathcal{D}_{train}, S_{in})\)
  • Not possible to condition only on some aspects of \(\mathcal{D}\).
• Normative: Counterfactual groups = stereotypes
  • Approach assumes attitudes are determined by traits.
  • Single answer imposes monolithic view for demographic subgroup.
• Santurkar et al. (2023) find their approach does not “work”:
  • Prompt does not make GPT return opinions representative of group

Innovation 3: GPT as Public Opinion

• P. Y. Wu et al. (2023) ask ChatGPT to choose the more liberal/conservative senator from given pairs.
• Apply Bradley-Terry model to estimate latent ideological score (ChatScores)
• Find that ChatScores better predict human evaluations than NOMINATE and CFscores.

Whose Opinions?

• Santurkar et al. (2023) compares answers from GPT to US public opinion in Pew Research poll.
• Finds substantial misalignment between views of LMs and public: equivalent to Dem-Rep divide on climate change.
• Instruction-tuning makes models even less representative.

Transparency, Reproducibility and Access

• GPT is closed-source and proprietary:
  • We don’t know the full extent of the training data.
  • We don’t know the exact architecture.
  • Hard to explain or predict behavior.
• Reproducibility:
  • Language generation can be deterministic, but usually not.
  • Prior versions of models may not be available in the future
• Access:
  • GPT is fairly affordable: text-davinci-003 (InstructGPT 175B, probably) is 0.02 USD/1000 tokens
  • But this can add up: applying a short zero-shot prompt to a corpus of 10k sentences costs 20 USD

Some Guidelines for Using GPT in Social Research

Do:

• Use it as a technical assistant (programming, how-to)
• Use it as a creative brainstorming tool (titles, pitches)

Open-Source

• Open Source LLMs exist
  • From HuggingFace, Meta, EleutherAI
  • Palmer and Spirling (2023) use OPT-30B (from Meta)
• Pros: auditable and reproducible
• Cons: massive hardware resource requirement

Smaller Models

• Alpaca (from Stanford CRFM): Instruction-tuned 7B parameter model
• Can zero-shot performance be “transferred” by generating synthetic labels?

Domain-Specific Models

• Domain-specific models may outperform larger ones (Kocoń et al. 2023)
• Bloomberg GPT (S. Wu et al. 2023)
• Ensembling LMs (Li et al. 2022; Gururangan et al. 2023)

Multimodal Models

• GPT-4 is image+text
• Audio, video

GPT-Easy

• Simple web-based interface for using GPT at scale
• Will have “guard rails” and transparent defaults built in
• Currently in development:
  • Looking for beta testers!

Topics I didn’t cover (and where to find it)

Technical:

• Transfer Learning
• Recursive Neural Networks and Sequence Modelling
• Decoder-only Transformers
• Encoder-Decoder Transformers
• Multi-task Learning

Task Learners

Turns out many tasks can be constructed as text completion:

Instruction-Tuning: In Words

1. Use human annotators to generate ideal responses to selection of prompts.
2. Use GPT-3 fine-tuned on human responses to generate multiple (synthetic) responses.
3. Use human annotators to rank synthetic responses.
4. Train a Reward Model on prompt+responses+ranking to emulate human scores.
5. Iteratively train GPT-3 with Reward Model and PPO.